January 07, 2026 Looking back at LLM security in 2025: the confusable deputy problem and deterministic boundaries